pi-no-autowrite

Pi extension: intercepts write/edit tool calls with two modes.

⚠️ Personal-use plugin. Published solely for multi-device sync and npm publishing experimentation. No guarantees of generality.

English | 中文

🇬🇧 English

Installation

pi install npm:pi-no-autowrite

Two Modes

| | education | supervisor | |---|---|---| | write/edit | Hard block, returns guidance for AI to teach you to hand-write | Allows, only shows a hint suggesting delegation to sub-agent | | system prompt | Unchanged | Injects supervisor instructions, telling AI it's a supervisor not an implementer | | mindset | You want to write code yourself, AI as teacher | You want AI to do the work but don't want context polluted by large code output |

Core difference: whether tool calls are actually blocked. Education blocks. Supervisor always allows — it only shifts AI behavior through system prompt injection.

Commands

/no-autowrite                  Toggle on/off
/no-autowrite mode <name>      Switch mode (education|supervisor), auto-enables
/no-autowrite on               Enable
/no-autowrite off              Disable
/no-autowrite status           Show status
/no-autowrite show             Show interception rules for current mode

Technical Implementation

Three core mechanisms:

1. tool_call hook — intercept write/edit

Registers pi.on("tool_call"), matching toolName === "write" || "edit".

• education: returns { block: true, reason: "..." }, blocking the call and instructing AI to guide the user instead
• supervisor: shows a notify hint then returns undefined, allowing the call

2. before_agent_start hook — inject system prompt

Active only in supervisor mode. Before AI starts, appends supervisor role definition to event.systemPrompt:

• Role: supervisor, not implementer
• Rules: heavy coding → delegate via Agent tool + run_in_background: true
• Light tasks (TODOs, comments, config, short snippets) → write/edit directly
• Use get_subagent_result to retrieve results and report back

This nudges AI toward delegation rather than direct implementation — but it's non-enforcing, just a soft prompt-level steer.

3. session entry — state persistence

Toggle state and mode are persisted via pi.appendEntry("no-autowrite-state", { mode, enabled }) into the session file, restored from the most recent entry on next startup. State survives project switches.

Motivation

Two conflicting needs when using pi daily:

• Sometimes you want to write code yourself — AI should only explain, not touch files
• Sometimes you want AI to go all out — but don't want the main session context flooded with code output

This plugin is the quick toggle between those two modes.

License

MIT

🇨🇳 中文

安装

pi install npm:pi-no-autowrite

两种模式

| | education | supervisor | |---|---|---| | write/edit | 硬拦截，返回引导信息让 AI 教你手写 | 放行，仅弹提示建议委托子 agent | | system prompt | 不改 | 注入 supervisor 行为指引，告诉 AI 它是监督者不是执行者 | | 适用心态 | 想自己写代码，AI 当老师 | 想让 AI 干活但不想 context 被大段代码污染 |

两者本质区别：是否真的阻止工具调用。education 直接 block，supervisor 永远放行——只是通过 system prompt 改变 AI 的行为倾向。

命令

/no-autowrite                  切换开关
/no-autowrite mode <name>      切换模式（education|supervisor），自动启用
/no-autowrite on               启用
/no-autowrite off              关闭
/no-autowrite status           查看状态
/no-autowrite show             查看当前拦截规则

技术实现

三个核心机制：

1. tool_call hook — 拦截 write/edit

注册 pi.on("tool_call")，匹配 toolName === "write" || "edit"。

• education：返回 { block: true, reason: "..." }，阻止本次调用，并提示 AI 改为指导用户
• supervisor：弹出 notify 提醒后返回 undefined，放行

2. before_agent_start hook — 注入 system prompt

仅 supervisor 模式生效。在 AI 启动前，向 event.systemPrompt 尾部追加 supervisor 角色定义：

• 角色：supervisor（监督者），不是 implementer
• 规则：繁重编码用 Agent tool + run_in_background: true 委托给后台子 agent
• 小东西（TODO、注释、配置、简短片段）可直接 write/edit
• 用 get_subagent_result 取回结果后汇报

这让 AI 在行为上更倾向委托而非亲自写——但不强制，只是一个 prompt 层面的软引导。

3. session entry — 状态持久化

开关状态和当前模式通过 pi.appendEntry("no-autowrite-state", { mode, enabled }) 写入 session，下次启动时从最近的 entry 恢复。切换项目后状态不丢失。

开发动机

日常使用 pi 时有两种矛盾需求：

• 有时想自己动手写代码，希望 AI 只解释思路不动文件
• 有时想让 AI 放开了干，但又不想主会话被代码输出撑爆 context

这个插件就是在这两种需求间切换的快捷键。

License

MIT